On Fri, May 30, 2014 at 08:15:43PM +0200, Samuel Thibault wrote: > Justus Winter, le Fri 30 May 2014 09:35:37 +0200, a écrit : > > Quoting Richard Braun (2014-05-29 19:12:13) > > > On Thu, May 29, 2014 at 07:04:48PM +0200, Samuel Thibault wrote: > > > > But precisely: once the only thread gets a data_request, it'll call > > > > pager_read_page, which will typically call store_read, which will wait > > > > on the eventual I/O, and no other request will be processed during that > > > > time, and thus not pipelining the I/O requests. > > > > > > Right, I assumed the store interface wasn't synchronous... > > > > So we want a fixed number of pager threads? > > We can use a fixed number of pager threads, yes. Perhaps just 1 indeed > currently gives the best performance because the device backends that we > currently use don't get benefit from piling up requests. But at least > we keep the infrastructure to pile up requests for when we fix the > backends.
Agreed. -- Richard Braun